AI Model Development

Feeds  Anomalies, FDD, GOC alarm wall, automatic work-order creation.

Consumes  15-minute point history per device (temperature, pressure, flow, kW, valve position, run status), equipment metadata including design conditions and capacity, weather.

{ siteId, equipmentId, faultClass, confidence, firstSeen, evidencePoints[], estimatedCostImpact }

Accepted when  Precision ≥ 0.85 against an engineer-labelled historical fault set, and every output cites the specific points that triggered it. An uncited fault cannot be displayed — the UI has no place to put it.

Feeds  Savings & ROI ledger, Client Portal savings claim, monthly client reports. This is the model that gets invoiced against, so it carries the most commercial risk of the ten.

Consumes  At least 12 months of pre-intervention interval consumption, weather (degree days), an occupancy or production driver, and the intervention log from the action ledger.

{ siteId, period, baselineKwh, actualKwh, avoidedKwh, confidenceInterval, cvrmse, nmbe, attribution[] }

Accepted when  IPMVP Option C is satisfied: CV(RMSE) ≤ 25% on monthly data and NMBE within ±0.5%. Outside those bounds the platform reports the saving as estimated rather than verified, and the model must return the statistics so the platform can make that call itself.

Feeds  Sensor Trust, and through it the baseline quality gate and the AI autonomy gate. Build this first: every other model trains and infers on data this one grades.

Consumes  Raw point streams, reference meter readings where a site has them, device metadata and expected reporting interval.

{ pointId, status: ok | frozen | drifting | gap | dead, driftPct, lastGoodAt, confidence }

Accepted when  Injected faults are detected within one hour: a frozen value, a 3% drift, and a 10% interval gap. Drift must be returned as null rather than zero where a site has no reference to compare against — the platform distinguishes unmeasured from good.

Feeds  Tariff & Demand. Indian HT tariffs set the demand charge from the single highest 15-minute average in the month, at 1.5× above the contracted ceiling, so a missed peak prediction has a direct rupee cost.

Consumes  15-minute kW history, weather forecast, calendar and occupancy, equipment start schedule.

{ siteId, horizon, forecastKw[], peakRiskWindow, breachProbability, confidence }

Accepted when  MAPE ≤ 8% at a four-hour horizon, and recall ≥ 0.9 on demand-ceiling breaches. Recall matters more than precision here: a false alarm costs an operator two minutes, a missed peak costs a month of penalty.

Feeds  AI action ledger. Runs in Approve mode by default: the model drafts, a human commits. Per-site autonomy is set in AI Governance and the model must honour it rather than assume it.

Consumes  Current plant state, M4 forecast, tariff structure, comfort and safety constraints, equipment performance curves.

{ siteId, actions[{ pointId, currentValue, proposedValue, holdUntil, revertTo }], expectedKwh, expectedRupees, rationale, confidence }

Accepted when  No proposal ever falls outside the comfort or safety envelope, every action carries a revert value, and realised savings reconcile with M2 after the fact. A recommendation whose saving cannot later be measured is not a recommendation.

Feeds  Global Operations Centre alarm wall.

Consumes  Live alarm stream with timestamps, equipment topology from the published SLD drawings, historical co-occurrence.

{ groupId, rootAlarmId, childAlarmIds[], causeHypothesis, confidence }

Accepted when  At least 70% of multi-alarm events collapse into one group, and no genuinely independent alarm is ever suppressed as a child. The second condition is absolute — a suppressed real alarm is a safety failure, not a quality metric.

Feeds  Alarm Rationalisation, measured against EEMUA 191.

Consumes  90 days of alarm history per point, equipment state, shelving history.

{ pointId, class: chattering | standing | duplicate | genuine, activations24h, standingDays, suggestedFix }

Accepted when  Agreement ≥ 0.8 with engineer classification on a labelled sample, and the suggested fix is specific enough to action — an on-delay value, a deadband width, a suppression condition, not "review this point".

Feeds  Field App work orders, planned maintenance scheduling.

Consumes  Runtime hours, vibration, filter differential pressure, motor current trends, service and replacement history.

{ equipmentId, failureMode, predictedWindow, confidence, recommendedAction, partsRequired[] }

Accepted when  Lead time is at least 7 days — shorter than that and a technician cannot be scheduled or a part ordered, which makes the prediction worthless — with a false-positive rate ≤ 20%.

Feeds  The recoverable-cost panel in Tariff & Demand: load shifting, staggered starts, power-factor correction, DG versus grid changeover.

Consumes  DISCOM tariff schedule per site (ToD bands, demand contract, PF incentive), M4 forecast, deferrable load inventory, DG fuel cost.

{ siteId, levers[{ action, expectedRupees, confidence, constraint }], blendedRate }

Accepted when  Modelled cost reconciles with the actual DISCOM bill within 5%. Until it does, the recommendations cannot be shown to a client, because the first thing they will do is compare.

Feeds  The AI Copilot tab. Retrieval over the registry, point history, action ledger and site documents.

Consumes  Natural-language question plus the retrieval corpus, scoped to the asking user's site permissions.

{ answer, citations[{ sourceType, sourceId, value, asOf }], confidence, refused }

Accepted when  Every number in an answer traces to a tag, record or document, and the model refuses rather than estimates when it cannot cite. Tenant scoping is enforced server-side, not by prompt.